A Poisson-based adaptive affinity propagation clustering for SAGE data

Abstract

Serial analysis of gene expression (SAGE) is a powerful tool to obtain gene expression profiles. Clustering analysis is a valuable technique for analyzing SAGE data. In this paper, we propose an adaptive clustering method for SAGE data analysis, namely, PoissonAPS. The method incorporates a novel clustering algorithm, Affinity Propagation (AP). While AP algorithm has demonstrated good performance on many different data sets, it also faces several limitations. PoissonAPS overcomes the limitations of AP using the clustering validation measure as a cost function of merging and splitting, and as a result, it can automatically cluster SAGE data without user-specified parameters. We evaluated PoissonAPS and compared its performance with other methods on several real life SAGE datasets. The experimental results show that PoissonAPS can produce meaningful and interpretable clusters for SAGE data.

Introduction

Serial analysis of gene expression (SAGE) is a powerful tool for obtaining global gene expression profiling at mRNA level. The original technique was provided by Dr. Victor Velculescu (Velculescu et al., 1995) in 1995. SAGE has been used to study the transcriptome of a variety of tissue and cell types from a diverse set of organisms (Zuyderduyn, 2007). Unlike microarray technique, SAGE detects unknown transcripts without requiring prior knowledge of what is present in the sample under analysis (Wang, 2007), and can provide a statistical description of the mRNA population present in a cell. Wang (Wang, 2007) presented a review of the features of the SAGE data, including the specificity of SAGE tags with respect to their original transcripts, the quantitative nature of SAGE data for differentially expressed genes, the reproducibility, the comparability of SAGE with microarray and the future potential of SAGE.

To analyze the significant amounts of SAGE data produced from this technology, a number of methods have been developed to deal with them. A SAGE data set can be viewed as an n × d matrix, where n is the number of tags and d is the number of SAGE libraries, and each horizontal row of matrix represents a SAGE tag and each vertical column represents different development stages of a biological process or various biological conditions (Wang et al., 2008). Clustering algorithms (Cai et al., 2004, Huang et al., 2008, Sander et al., 2005, Tzanis and Vlahava, 2007, Wang et al., 2007, Wang et al., 2008) provide a useful tool to explore the potentially novel and significant transcript or gene groups in SAGE data. The basic concept of clustering is to divide patterns into different groups (clusters). The patterns in the same share more similarity comparing with the patterns in other clusters. A nice review on the clustering methods for SAGE data can be found in Wang et al. (2008). As described earlier, unlike microarray expression technology, SAGE produces profiles consisting of a digital output that is quantitative in nature. Therefore, traditional distances or similarity measures, such as Pearson correlation coefficient and Euclidean distance, may not be suitable for SAGE data analysis. Evidently, clustering method for SAGE data should employ a reliable statistical model (Cai et al., 2004, Lu et al., 2005, Wang et al., 2008, Zuyderduyn, 2007). For this purpose, Cai et al. (Cai et al., 2004, Huang et al., 2008) modeled SAGE data by Poisson statistics and developed two Poisson-based distances, and implemented the k-means clustering using two Poisson-based distances. Afterwards, Thygesen and Zwinderman (2006) proposed a hierarchical Poisson model with a gamma prior and three different algorithms for estimating the parameters in the model. Furthermore, Wang et al. (2007) proposed two new clustering algorithms, namely, PoissonS and PoissonHC based on the adaptation and improvement of Self-Organizing Maps (SOM) and hierarchical clustering techniques. Lu et al. (2005) employed an overdispersed log-linear model approach to analyze SAGE libraries. Vencio et al. (2004) proposed a Bayesian model of mixtures to account for within-class variability.

Although a number of clustering methods for SAGE data have been proposed (Cai et al., 2004, Huang et al., 2008, Sander et al., 2005, Tzanis and Vlahava, 2007, Wang et al., 2007, Wang et al., 2008), most of these methods employ some user-defined parameters, therefore the results may highly depend on such parameters. For example, k-means and k-medoids face the problem of prescribing the number of clusters in advance. In addition, many clustering algorithms for SAGE data start with a randomly initial selection, such as k-means and PoissonC. Consequently, the clustering result cannot be reproduced. Clearly, except for specific situations when we have complete knowledge about the data set to ensure the validity of chosen parameters, the choice of the parameters can only be determined by empirical methods.

Recently, Frey and Dueck (2007) proposed a powerful algorithm named Affinity Propagation (AP) based on message passing. AP algorithm has attracted increasing attention. However, the question how selection suitable parameters should be made has received only little attention in the original literature. Based on AP and Poisson statistics, this paper proposes an adaptive clustering method for SAGE data analysis using clustering validation measure as a cost function of merging and splitting, namely, PoissonAPS. The key characteristics of the proposed methods are as follows: (1) the method overcomes some limitations of AP; (2) the method is a non-parametric clustering method.

The organization of this paper is as follows. In Section 2, we give a brief review of AP clustering algorithm and statistical properties of SAGE data. Then, we introduce the proposed method in Section 3. Section 4 presents the experimental results. Discussions and conclusions are given in Section 5.